Product packaging parameter optimization method and system based on data joint analysis

Abstract

The application discloses a finished product packaging parameter optimization method and system based on data joint analysis, and belongs to the technical field of image processing. The method obtains real-time packaging execution state data, packaging visual image data, upstream incoming material state data and downstream logistics disturbance data of the finished product packaging in a current execution cycle; aligns the above-mentioned multi-source data in a time dimension to generate a multi-source joint analysis data set; inputs into a pre-constructed joint analysis model to output a packaging parameter compensation adjustment value of a next execution cycle; updates a target packaging parameter according to the compensation adjustment value, and controls a packaging execution mechanism to operate. The joint analysis model is used for describing the correlation between the packaging execution state, the incoming material state, the logistics disturbance and the appearance representation after packaging. Through the joint analysis of the visual image data and the multi-source data of the upstream and downstream of the production line, the application realizes the optimization of the packaging parameter from passive correction to active pre-adjustment, and improves the consistency of the finished product packaging quality.

Description

Technical Field

[0001] This application relates to the field of image processing technology, specifically to a method and system for optimizing finished product packaging parameters based on joint data analysis. Background Technology

[0002] Currently, in the industrial finished product packaging field, packaging parameter settings typically rely on fixed process standards of the production line or feedback control based on a single sensor. For example, existing technologies often use visual inspection equipment to identify surface defects in packaged products, such as the sealing position and whether the label is crooked. When defects exceed the standard, an alarm is triggered or the machine is stopped, requiring manual intervention to adjust parameters such as heat sealing temperature, pressure, or labeling position. This approach is essentially a "post-mortem correction," failing to predict and optimize parameters before defects occur.

[0003] Analysis reveals that existing methods are limited by their singular data sources and lagging analytical logic. The final quality of packaging depends not only on the instantaneous equipment status (e.g., temperature, pressure) but also on the dynamic disturbances of preceding processes (e.g., finished product filling volume, material characteristics) and subsequent processes (e.g., conveying, palletizing). For example, minute fluctuations in filling volume can cause changes in pressure within the packaging bag, thus affecting the optimal heat-sealing temperature; while misalignment between conveyor belt vibration frequency and labeling timing can lead to label wrinkles. Existing technologies fail to effectively integrate and jointly analyze these real-time data from different sources and dimensions on the production line, resulting in a lack of holistic and predictive capabilities in optimizing packaging parameters.

[0004] Therefore, there is an urgent need for a solution based on joint data analysis to optimize finished product packaging parameters. Summary of the Invention

[0005] This application provides a method and system for optimizing finished product packaging parameters based on joint data analysis, which at least addresses the problems existing in the prior art.

[0006] A first aspect of this application provides a method for optimizing finished product packaging parameters based on joint data analysis, comprising the following steps: S1: Obtain real-time packaging execution status data of the finished product packaging within the current execution cycle; S2: After the current execution cycle ends, collect the visual image data of the finished product packaging after packaging; S3: Obtain upstream incoming material status data and downstream logistics disturbance data corresponding to finished product packaging; S4: Align real-time packaging execution status data, post-packaging visual image data, upstream material arrival status data, and downstream logistics disturbance data in the time dimension to generate a multi-source joint analysis dataset. S5: Input the multi-source joint analysis dataset into the pre-built joint analysis model and output the packaging parameter compensation adjustment value for the next execution cycle; where the joint analysis model is a model used to describe the relationship between packaging execution status, incoming material status, logistics disturbance and post-packaging appearance characterization; S6: Update the target packaging parameters for the next execution cycle based on the packaging parameter compensation adjustment value; S7: In the next execution cycle, control the operation of the packaging actuator based on the updated target packaging parameters.

[0007] This application embodiment acquires real-time packaging execution status data, post-packaging visual image data, upstream material arrival status data, and downstream logistics disturbance data for the current execution cycle. After aligning these multi-source data along the time dimension, they are input into a joint analysis model, enabling the output of packaging parameter compensation adjustment values ​​for the next execution cycle. This approach helps to optimize packaging parameters from passive correction to proactive pre-adjustment, allowing the packaging process to adapt to fluctuations in upstream materials and downstream environmental disturbances.

[0008] In some embodiments of this application, the multi-source joint analysis dataset is input into a pre-built joint analysis model, and the packaging parameter compensation adjustment value for the next execution cycle is output. This includes: extracting features from the multi-source joint analysis dataset through the joint analysis model to obtain a joint feature vector; matching the joint feature vector with a preset standard packaging feature template to determine the type and degree of packaging quality deviation in the current execution cycle; and calculating the packaging parameter compensation adjustment value to offset the deviation based on the deviation type and degree.

[0009] This application's embodiments extract joint feature vectors from multi-source data using a joint analysis model and match them with standard packaging feature templates to determine the type and degree of packaging quality deviation in the current execution cycle. Based on the deviation type and degree, compensation adjustment values ​​are calculated, enabling more targeted mitigation of specific deviation factors causing packaging quality problems.

[0010] In some embodiments of this application, the joint analysis model is constructed using a deep neural network. The input layer of the deep neural network includes multiple feature input branches corresponding to real-time packaging execution status data, post-packaging visual image data, upstream incoming material status data, and downstream logistics disturbance data, respectively. The multiple feature input branches are fused in the intermediate layer of the network to generate a joint feature vector.

[0011] This application embodiment constructs a joint analysis model using a deep neural network with multiple feature input branches. Different branches process real-time packaging execution status data, post-packaging visual image data, upstream incoming material status data, and downstream logistics disturbance data, respectively, and feature fusion is performed in the intermediate layer. This network structure can better learn the complex relationships between multi-source heterogeneous data, which helps to improve the representational ability of the joint feature vector.

[0012] In some embodiments of this application, the acquisition of upstream incoming material status data and downstream logistics disturbance data corresponding to the finished packaging includes: acquiring the real-time filling weight and filling density when filling to form the contents of the finished packaging, as upstream incoming material status data; and acquiring the real-time operating speed and vibration amplitude of the conveying mechanism during the process of the finished packaging being transported to the next workstation after packaging, as downstream logistics disturbance data.

[0013] This application's embodiments acquire upstream incoming material status data, including real-time filling weight and density during filling, and downstream logistics disturbance data, including real-time operating speed and vibration amplitude of the conveyor mechanism. This data reflects the physical characteristics of the packaged contents and the dynamic environment of the post-packaging conveying process, providing more comprehensive input information for joint analysis.

[0014] In some embodiments of this application, real-time packaging execution status data, post-packaging visual image data, upstream material arrival status data, and downstream logistics disturbance data are aligned in the time dimension, including: taking the end time of the current execution cycle as a benchmark, tracing back to the upstream material filling time corresponding to the finished product packaging, associating the upstream material arrival status data generated at the upstream material filling time with the end time; and binding the real-time packaging execution status data and post-packaging visual image data corresponding to the end time with the associated upstream material arrival status data using timestamps.

[0015] This embodiment of the application uses the end time of the current execution cycle as a baseline, traces back to the upstream material filling time, associates the data generated at the upstream material filling time with the end time, and then timestamps the packaging execution status data and post-packaging visual image data corresponding to the end time with the associated upstream material status data. This alignment method can establish the causal relationship between the data of each stage of the packaging process in time, providing a time-consistent data foundation for subsequent analysis.

[0016] In some embodiments of this application, the real-time packaging execution status data includes at least the heat sealing temperature, heat sealing pressure, and labeling position coordinates.

[0017] This application embodiment sets real-time packaging execution status data, including heat-sealing temperature, heat-sealing pressure, and labeling position coordinates. These parameters are key process variables affecting packaging sealing quality and label placement; obtaining this data helps analyze the extent to which the packaging execution process affects the final packaging quality.

[0018] In some embodiments of this application, the visual image data after packaging includes at least a texture feature map of the sealed area, a measurement of the sealed width, and the coordinates of the corner positions of the label.

[0019] This application embodiment sets the visual image data after packaging to include the texture feature map of the sealing area, the measured value of the sealing width, and the coordinates of the corner positions of the label. These image features can quantitatively reflect the integrity of the packaging seal and the accuracy of the label affixing, providing a visual and objective basis for judging the packaging quality.

[0020] In some embodiments of this application, before inputting the multi-source joint analysis dataset into the pre-built joint analysis model, the method further includes: acquiring historical real-time packaging execution status data, historical post-packaging visual image data, historical upstream incoming material status data, historical downstream logistics disturbance data, and their corresponding historical compensation adjustment values ​​within a historical production cycle; using the historical real-time packaging execution status data, historical post-packaging visual image data, historical upstream incoming material status data, and historical downstream logistics disturbance data as training samples, and using the corresponding historical compensation adjustment values ​​as supervision labels, training the initial neural network model to obtain the joint analysis model.

[0021] This application embodiment trains an initial neural network model by acquiring multi-source historical data and corresponding historical compensation adjustment values ​​from historical production cycles before applying the joint analysis model. The historical data serves as the training sample, and the historical compensation adjustment values ​​are used as the supervision label. This method of training the model based on historical experience data enables the model to learn the patterns of packaging parameter adjustments under different operating conditions.

[0022] In some embodiments of this application, the historical compensation adjustment value is based on the historical best adjustment experience value marked manually or the best adjustment value obtained through closed-loop control test calibration.

[0023] In this embodiment, the historical compensation adjustment value is set based on the historical best adjustment experience value marked by manual annotation or the best adjustment value obtained through closed-loop control experiment calibration. Both of these adjustment values ​​are optimized values ​​that have been verified in practice or determined by experiments. Using these as training labels helps to improve the output accuracy of the joint analysis model.

[0024] In some embodiments of this application, updating the target packaging parameters for the next execution cycle based on the packaging parameter compensation adjustment value can be achieved in the following way: First, obtain the initial target packaging parameters preset for the next execution cycle. These initial target packaging parameters are baseline values ​​predetermined based on production process standards or historical statistical data. Then, the packaging parameter compensation adjustment value output from the joint analysis model is superimposed with the initial target packaging parameters. The specific method of superposition calculation can be addition, that is, adding the values ​​of each corresponding parameter dimension separately. For example, adding the initial target heat sealing temperature with the heat sealing temperature compensation adjustment value, adding the initial target heat sealing pressure with the heat sealing pressure compensation adjustment value, and adding the initial target labeling position coordinates with the labeling position compensation adjustment value. After superposition calculation, a set of updated target packaging parameters is generated. This set of updated parameters includes the baseline process requirements and the deviation compensation requirements identified by the joint analysis results of multi-source data from the previous execution cycle. The updated target packaging parameters are then sent to the control system of the packaging execution mechanism as the control target for the actual operation of the next execution cycle.

[0025] This application's embodiment generates updated target packaging parameters by superimposing initial target packaging parameters with packaging parameter compensation adjustment values. This approach, while maintaining the stability of the original process baseline, incorporates dynamic compensation requirements derived from multi-source data joint analysis. This ensures that the parameter adjustment process does not deviate excessively from the original process specifications, and can appropriately correct for identified deviation factors, thus helping to improve the accuracy and stability of parameter adjustment.

[0026] In some embodiments of this application, the method may further include a model incremental update process based on positive samples. After the next execution cycle ends, updated visual image data of the finished product packaging produced in that cycle is acquired through a visual acquisition device; the updated visual image data is input to a quality evaluation module, which performs comparative analysis based on preset appearance quality standards; the preset appearance quality standards may include indicators such as sealing integrity threshold, label position tolerance range, and allowable deviation of sealing width; if all indicators of the updated visual image data are within the preset standard range, the packaging quality of that execution cycle is determined to meet the requirements; at this time, the packaging parameter compensation adjustment value used in that execution cycle and the multi-source joint analysis dataset used to generate the compensation adjustment value are extracted and combined to form a positive sample; this positive sample is added to the training sample library of the joint analysis model for incremental training of the model.

[0027] This embodiment of the application acquires updated visual image data after packaging at the end of the next execution cycle. If the image data meets preset standards, the corresponding packaging parameter compensation adjustment values ​​and their multi-source joint analysis dataset are used as positive samples for incremental training of the joint analysis model. This approach enables the model to continuously learn from successful adjustments, gradually improving the accuracy of the output compensation adjustment values, and helping the model adapt to slow changes in production line conditions or fluctuations in material properties.

[0028] In some embodiments of this application, the finished product packaging can adopt a pre-made bag filling packaging form using flexible packaging materials; flexible packaging materials include plastic film, aluminum foil composite film, or composite materials of paper and plastic; pre-made bags refer to pre-formed and opened packaging bags, in which the contents are filled into the bag by a filling mechanism during the filling process, and then the bag enters the heat sealing process to complete the sealing; the packaging execution mechanism may include a film pulling mechanism, a heat sealing mechanism, and a labeling mechanism; the film pulling mechanism is used to pull the film material forward in a continuous packaging machine and control the bag length and positioning accuracy of each packaging unit; the heat sealing mechanism includes a transverse heat sealing assembly and a longitudinal heat sealing assembly, which, through the cooperation of a heating element and a pressurizing cylinder, applies heat and pressure to the bag opening at a set temperature to complete the fusion seal; the labeling mechanism is located after the heat sealing process or in parallel with the heat sealing process, and affixes the label to the designated position of the packaging bag.

[0029] The embodiments of this application optimize parameters such as heat sealing temperature, heat sealing pressure, and labeling position, which helps to reduce problems such as poor sealing or label misalignment caused by fluctuations in upstream incoming materials and disturbances in downstream conveying.

[0030] A second aspect of this application provides a finished product packaging parameter optimization system based on data joint analysis. This system is used to implement the aforementioned finished product packaging parameter optimization method based on data joint analysis, including: The first data acquisition module is used to acquire real-time packaging execution status data of finished product packaging within the current execution cycle; The second data acquisition module is used to collect visual image data of the finished product packaging after the current execution cycle ends. The third data acquisition module is used to acquire upstream material status data and downstream logistics disturbance data corresponding to the finished product packaging. The data alignment module is used to align real-time packaging execution status data, post-packaging visual image data, upstream material arrival status data, and downstream logistics disturbance data in the time dimension to generate a multi-source joint analysis dataset. The compensation value calculation module is used to input the multi-source joint analysis dataset into the pre-built joint analysis model and output the packaging parameter compensation adjustment value for the next execution cycle; wherein, the joint analysis model is a model used to describe the relationship between packaging execution status, incoming material status, logistics disturbance and post-packaging appearance; The parameter update module is used to update the target packaging parameters for the next execution cycle based on the packaging parameter compensation adjustment value. The execution control module is used to control the operation of the packaging execution mechanism based on the updated target packaging parameters in the next execution cycle.

[0031] In the above embodiments, the packaging parameter optimization process can be automated and made real-time through the collaborative work of each module. Attached Figure Description

[0032] Figure 1 A flowchart illustrating a method for optimizing finished product packaging parameters based on joint data analysis, provided as an embodiment of this application; Figure 2 A schematic diagram of a finished product packaging parameter optimization system based on joint data analysis is provided in one embodiment of this application; Figure 3 This is a schematic diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0033] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0034] To make the purpose, technical solution, and advantages of this application clearer, the following will be described in conjunction with the appendix. Figure 1-3 The following is an explanation using specific examples.

[0035] Please refer to Figure 1 , Figure 1 An embodiment of this application provides a method for optimizing finished product packaging parameters based on joint data analysis, comprising the following steps: S1: Obtain real-time packaging execution status data of the finished product packaging within the current execution cycle; S2: After the current execution cycle ends, collect the visual image data of the finished product packaging after packaging; S3: Obtain upstream incoming material status data and downstream logistics disturbance data corresponding to finished product packaging; S4: Align real-time packaging execution status data, post-packaging visual image data, upstream material arrival status data, and downstream logistics disturbance data in the time dimension to generate a multi-source joint analysis dataset. S5: Input the multi-source joint analysis dataset into the pre-built joint analysis model and output the packaging parameter compensation adjustment value for the next execution cycle; where the joint analysis model is a model used to describe the relationship between packaging execution status, incoming material status, logistics disturbance and post-packaging appearance characterization; S6: Update the target packaging parameters for the next execution cycle based on the packaging parameter compensation adjustment value; S7: In the next execution cycle, control the operation of the packaging actuator based on the updated target packaging parameters.

[0036] Currently, the setting and adjustment of packaging parameters on industrial packaging production lines mainly rely on two modes. One is the fixed parameter mode, where parameters such as heat sealing temperature, pressure, and labeling position are pre-set according to the packaging material specifications and product type, and the production line operates according to these fixed parameters long-term. The other is the post-production feedback mode, which uses online visual inspection equipment to perform appearance inspection on the packaged products. When defects such as incomplete sealing, misaligned labels, or wrinkles are detected and exceed the set threshold, the system issues an alarm signal, requiring operators to stop the machine, check, and manually adjust the relevant parameters, or the control system triggers a preset parameter correction scheme based on a single defect type. Both modes have their limitations in practical applications.

[0037] A thorough analysis of the existing models reveals that their root causes lie in the one-sided use of data and the lag in decision-making logic. Fixed-parameter models cannot handle fluctuations in upstream materials. For example, the filling amount of material in the same batch at different times may vary slightly due to pressure fluctuations in the supply system. This difference leads to changes in pressure inside the packaging bag, causing the originally appropriate heat-sealing temperature to become too high or too low. While post-feedback models introduce visual inspection, the results are only used to determine whether the current product is qualified, without establishing a correlation with factors such as process parameters, incoming material status, and conveying environment that caused the defects. For instance, a deviation in label placement might be caused by positioning misalignment of the labeling mechanism itself, changes in bag size due to film shrinkage during heat sealing in the previous process, or instability in the bag's position upon arrival at the labeling station due to conveyor belt vibration. Due to the lack of joint analysis of this multi-source data, operators find it difficult to accurately determine the true cause of the deviation, often relying on trial and error when adjusting parameters, resulting in low efficiency and inconsistent effectiveness.

[0038] This application, through systematic observation of the packaging process, reveals that the final packaging quality is the result of multiple coupled factors. The equipment status at the moment of packaging execution (e.g., heat-sealing temperature, pressure) directly affects the packaging material, determining the fusion quality of the sealing area. The appearance characteristics after packaging (e.g., sealing texture, label position) are a direct reflection of this effect. Simultaneously, the upstream material condition (e.g., filling weight, filling density) determines the internal support force of the packaging contents on the bag, affecting the material stress state in the heat-sealing area. Downstream material disturbances (e.g., conveyor speed, vibration) affect the positional accuracy of the packaging at subsequent workstations, particularly significantly impacting labeling accuracy. Based on this understanding, a framework for joint analysis of these four types of data can be constructed. First, execution status data is collected in real time; visual image data is acquired after packaging completion; and upstream material data and downstream disturbance data corresponding to the packaging are traced. Then, these data are aligned along the time dimension to form a joint dataset that comprehensively describes the entire process of the packaging from filling to conveying. Finally, a pre-built joint analysis model is used to identify the correlation patterns between execution status, incoming material status, disturbance status, and final appearance characterization from this data. Based on the deviation trend identified in the current cycle, the compensation adjustment value to be applied in the next cycle is calculated. This approach transforms the optimization of packaging parameters from passively waiting for corrections after defects occur to proactively adjusting based on multi-source data to predict deviation trends.

[0039] Real-time packaging execution status data refers to the process parameter data collected in real time by sensors installed on various execution components during the packaging execution process of the current finished product packaging. This data reflects the specific operating status of the equipment at the moment of packaging execution. For example, it may include the actual heating temperature during heat sealing, the actual pressure value applied by the pressurizing mechanism, and the actual position coordinates of the labeling head when releasing the label.

[0040] In this embodiment, post-packaging visual image data refers to image information captured by industrial or smart cameras deployed on the production line after the packaging action in the current execution cycle is completed. These images, after processing, can extract various features reflecting the packaging's appearance quality, such as the texture clarity of the sealing area, the width of the sealing edge, and the actual position of the four corners of the label on the packaging surface. Upstream material status data refers to the material status data generated during the filling or forming process before the current finished packaging enters the packaging process. Since the filling of the packaging contents usually precedes the sealing process, this data needs to be traced and correlated according to the production cycle. For example, it may include the real-time filling weight, real-time density, or viscosity of the filling material during filling to form the packaging contents. Downstream logistics disturbance data refers to the dynamic environmental data generated by the conveying mechanism during the process of transporting the finished packaging to subsequent workstations after the packaging action is completed. This data reflects the mechanical disturbance experienced by the packaging during transport, which may affect the accuracy of subsequent inspection or labeling processes. For example, it may include the real-time running speed of the conveyor belt, the vibration amplitude generated by the conveyor platform, and the timing of the steering mechanism's action. Alignment along the time dimension refers to establishing a correlation between multi-source data from different processes and collection times, based on their correspondence with the same finished product packaging, on the timeline. Since upstream material arrival, packaging execution, post-packaging visual acquisition, and downstream transportation occur at different points in time, it is necessary to use the end time of the current execution cycle or the visual acquisition time as a baseline, tracing back to the upstream material arrival time corresponding to the packaging body, and then associating it with the downstream transportation time. This ensures that each record in the final multi-source joint analysis dataset completely describes the entire process status information of a specific packaging body from upstream material arrival to downstream transportation. The joint analysis model is a pre-trained computational model whose input is the multi-source joint analysis dataset, and whose output is the packaging parameter compensation adjustment value for the next execution cycle. This model establishes a complex mapping relationship between packaging execution status, upstream material arrival status, downstream logistics disturbance status, and the post-packaging visual appearance representation by learning from a large amount of historical data. When multi-source data for the current cycle is input, the model can identify the deviation trend existing in the current cycle and calculate the direction and magnitude of parameter adjustments needed to offset this deviation trend and make the packaging appearance in the next cycle closer to the ideal state. Packaging parameter compensation adjustment values ​​are corrections to the target packaging parameters for the next execution cycle. They do not directly replace the original target parameters, but rather add a dynamic adjustment to the original target parameters. For example, increasing the originally set heat-sealing temperature by 2 degrees Celsius, or shifting the labeling position coordinates by 0.5 millimeters in a certain direction. Updating the target packaging parameters for the next execution cycle means combining the calculated compensation adjustment values ​​with the initial target parameters originally planned for the next cycle to form a new control target.This new control target is then sent to the controller of the packaging actuator as the setpoint for the actual operation in the next cycle. The packaging actuator is a mechanical device that performs specific actions such as sealing and labeling of packaging, and may include heating elements that provide heat, pressure cylinders or servo motors that apply pressure, labeling heads that place labels, and film-pulling rollers that pull the film.

[0041] This technical solution constructs a complete closed loop from data acquisition to parameter adjustment and execution control through seven steps. The first step acquires real-time packaging execution status data, capturing equipment factors directly affecting material deformation during the packaging process. The second step acquires visual image data after packaging, quantifying the intuitive appearance of the packaging execution result. The third step acquires upstream material status data and downstream logistics disturbance data, introducing environmental influencing factors in both directions before and after the packaging process. These three data acquisition steps collect information from four dimensions: execution process, execution result, upstream influence, and downstream influence, providing a comprehensive data foundation for subsequent analysis. The fourth step, time dimension alignment, is a crucial step in data fusion. It solves the problem of difficulty in directly correlating multi-source data due to different acquisition times, ensuring that the final multi-source joint analysis dataset truly reflects the complete state of a package from upstream material arrival to downstream transportation, providing time-consistent and causally clear input for subsequent model analysis. The fifth step processes this multi-source joint analysis dataset through a joint analysis model, outputting compensation adjustment values. This step maps multi-dimensional state information into specific parameter correction instructions, realizing the transformation from state perception to decision output. The model's role is to identify the dominant factors causing appearance deviations from complex data relationships and calculate the parameter adjustments needed to offset these deviations. The sixth step updates the target packaging parameters for the next cycle based on the compensation adjustment values, transforming the model's decision into executable process settings. The seventh step controls the packaging actuator based on the updated target parameters, applying the adjusted parameters to the actual production process and completing the closed loop from decision-making to execution.

[0042] These seven steps are interconnected, forming a complete optimization cycle. Each cycle is based on multi-source data collected in the current period. After analysis and decision-making, the process parameters for the next cycle are adjusted, ensuring that the packaging parameters can continuously respond to fluctuations in upstream materials and changes in the downstream environment, gradually approaching the optimal state under the current operating conditions. Compared to the traditional fixed parameter mode, this method increases adaptability to changes in operating conditions. Compared to the post-event feedback mode, this method advances the adjustment time from correction after defects occur to proactive pre-adjustment based on trend prediction, helping to reduce the generation of defective products.

[0043] Taking a cigarette carton packaging production line as an example, the packaging material uses pre-made BOPP film pre-fabricated bags for packaging single cartons of cigarettes. In the current execution cycle, the control system collects real-time execution status data at a heat-sealing temperature of 142 degrees Celsius and a heat-sealing pressure of 0.3 MPa. After packaging, an industrial camera captures an image of the carton packaging. Processing reveals slight wrinkles and textures on the sealing edge, and the upper left corner of the label is 0.8 mm higher than the standard position. Simultaneously, the system traces upstream data and finds that the filling weight of the corresponding cigarette carton at the filling moment was 245 grams, slightly higher than the standard 240 grams. Downstream conveyor data records show that when the carton passed through the conveyor belt, the conveying speed was 0.5 meters per second, and the vibration sensor captured a vibration with an amplitude of 0.2 mm. The data alignment module combines this information into a multi-source joint analysis dataset. After processing this dataset, the joint analysis model outputs compensation adjustment values: the heat-sealing temperature is reduced by 3 degrees Celsius, and the label position's Y-axis coordinate is shifted down by 0.3 mm. The control system adjusts the initial target parameter for the next cycle, the heat sealing temperature, from 143 degrees Celsius to 140 degrees Celsius, and the labeling position is shifted downwards by 0.3 millimeters from the original target. In the next execution cycle, the heat sealing mechanism operates at 140 degrees Celsius, and the labeling mechanism applies the label at the adjusted coordinate position, thus improving the sealing wrinkles and making the label position closer to the standard position.

[0044] In some embodiments of this application, the multi-source joint analysis dataset is input into a pre-built joint analysis model, and the packaging parameter compensation adjustment value for the next execution cycle is output. This includes: extracting features from the multi-source joint analysis dataset through the joint analysis model to obtain a joint feature vector; matching the joint feature vector with a preset standard packaging feature template to determine the type and degree of packaging quality deviation in the current execution cycle; and calculating the packaging parameter compensation adjustment value to offset the deviation based on the deviation type and degree.

[0045] In this embodiment, feature extraction refers to the process by which the joint analysis model processes the input multi-source joint analysis dataset to extract high-dimensional feature information that reflects the essential attributes of the data. The input multi-source joint analysis dataset includes real-time packaging execution status data, post-packaging visual image data, upstream material arrival status data, and downstream logistics disturbance data. These data are diverse in form, some being numerical time-series data, and others being image data. The role of feature extraction is to convert these raw data into a unified feature vector, facilitating subsequent matching and calculation. For example, for post-packaging visual image data, feature extraction can identify visual features such as the texture direction of the sealing area, wrinkle density, and the straightness of the label edges; for real-time packaging execution status data, feature extraction can retain statistical features such as the trend of heat-sealing temperature changes and the amplitude of pressure fluctuations.

[0046] The joint feature vector (JV) is a multi-dimensional vector representation generated by the joint analysis model after extracting features from a multi-source joint analysis dataset. This vector integrates raw information from different data sources and with different physical meanings into a unified mathematical expression, with each dimension corresponding to an abstract feature learned by the model. The number of dimensions in the JV is determined by the model structure; for example, it can be a 128-dimensional or 256-dimensional vector. This vector condenses the state information of the entire process of finished product packaging in the current execution cycle, from upstream material arrival to downstream transportation, providing a basis for subsequent comparison with a standard template.

[0047] A pre-built standard packaging feature template is a pre-constructed baseline feature vector or set of vectors, representing the feature distribution that a multi-source joint analysis dataset should possess under ideal process conditions and without significant deviations. The standard packaging feature template can be obtained by collecting a large amount of multi-source data corresponding to qualified finished packaging, and then clustering or averaging the data after feature extraction. Different standard templates can be set for different specifications of finished packaging. For example, for a certain specification of cigarette box packaging, the joint feature vector corresponding to its standard template reflects the ideal state of clear sealing texture and centered label position under conditions such as a heat-sealing temperature of 145 degrees Celsius, a filling weight of 240 grams, and a conveying speed of 0.5 meters per second.

[0048] Packaging quality deviation type refers to the category of appearance quality problems that exist in the finished packaging of the current execution cycle relative to the preset standard. The joint analysis model compares the joint feature vector generated in the current cycle with the standard packaging feature template to identify the direction and pattern of difference between the two in the feature space, thereby determining the specific type of problem existing in the current packaging. Deviation types can include various categories such as loose sealing due to sealing temperature deviation, sealing wrinkles due to sealing pressure deviation, bag deformation due to fluctuations in incoming material filling, and label skew due to conveyor vibration.

[0049] Packaging quality deviation is a parameter that quantifies the type of deviation, reflecting the severity of the current packaging's deviation from the ideal state. The degree of deviation can be represented numerically, such as the area ratio of the sealing wrinkled region, the distance in millimeters of label offset, and the difference between the sealing width and the standard value. The degree of deviation can be determined based on the feature distance between the joint feature vector and the standard template, or it can be directly output by the model as a specific deviation value.

[0050] Calculating packaging parameter compensation adjustment values ​​to offset deviations refers to determining, based on the identified deviation type and degree, which packaging parameters should be adjusted in what way to make the packaging appearance in the next execution cycle closer to the ideal state, using the model's built-in mapping relationship or preset adjustment rules. Different deviation types correspond to different parameter adjustment strategies. For example, if the deviation type is a weak seal due to a low sealing temperature, the compensation adjustment value can be a positive increase in the heat-sealing temperature; if the deviation type is increased bag pressure and over-melting of the seal due to excessive filling weight, the compensation adjustment value can be a negative decrease in the heat-sealing temperature. The specific magnitude of the compensation adjustment value is related to the degree of deviation; the greater the degree of deviation, the larger the magnitude of the compensation adjustment value.

[0051] Understandably, the internal structure of a joint analysis model can be implemented in various ways. For example, a neural network-based encoder structure can be used, where the encoder is responsible for mapping the multi-source joint analysis dataset to a joint feature vector, and the decoder or classifier is responsible for outputting the bias type, bias degree, and compensation adjustment value based on the joint feature vector. Alternatively, an end-to-end learning approach can be used, directly mapping the joint feature vector to the compensation adjustment value.

[0052] This embodiment constructs a processing chain within the joint analysis model, from feature extraction to deviation identification and compensation calculation. First, feature extraction transforms diverse and meaningful multi-source raw data into a unified joint feature vector. This step solves the problem of direct comparison and computation of heterogeneous multi-source data, providing a unified input format for subsequent processing. The joint feature vector condenses the state information of the entire process from upstream material arrival to downstream conveying in the current execution cycle, encompassing a comprehensive expression of various factors affecting packaging appearance. Then, the joint feature vector is matched with a preset standard packaging feature template. By comparing the differences between the current state and the ideal state in the feature space, the deviation type and degree are determined. This step enables refined diagnosis of packaging quality problems, not only determining whether the product is qualified but also clarifying the specific type and severity of the problem. Determining the deviation type provides a basis for selecting the adjustment direction, while determining the deviation degree provides a quantitative reference for the adjustment range. Finally, based on the deviation type and degree, packaging parameter compensation adjustment values ​​are calculated, transforming the diagnostic results into specific execution instructions. Different types of deviations correspond to different parameter adjustment strategies, and different degrees of deviation correspond to different adjustment ranges, making the output compensation adjustment values ​​targeted and adaptable.

[0053] These three steps are interconnected, forming a complete logic from data compression to pattern recognition to decision output. Feature extraction provides effective input for matching, the matching results provide the basis for calculation, and the calculation process transforms the diagnostic results into actionable adjustments. Compared to the end-to-end approach of directly mapping multi-source joint analysis datasets to compensation adjustment values, this decomposition method of first identifying the type and degree of deviation and then calculating the compensation value makes the model's decision-making process more transparent and controllable. The intermediate outputs of deviation type and degree can also be used for monitoring and recording the production process, making it easier for operators to understand the current system status and the basis for adjustments.

[0054] Taking a cigarette box packaging production line as an example, after an execution cycle, the joint analysis model receives a multi-source joint analysis dataset containing data such as a heat-sealing temperature of 142 degrees Celsius, a heat-sealing pressure of 0.3 MPa, slight wrinkles in the sealing image, a filling weight of 245 grams, and a conveyor vibration of 0.2 mm. The model first extracts features through the encoder, generating a 128-dimensional joint feature vector. This vector is then compared with a pre-stored standard packaging feature template for this specification of cigarette box. The model finds that the feature vector deviates from the standard template in several dimensions representing the uniformity of the sealing texture, and also deviates slightly in the dimension representing the stability of the label position. The model thus determines that there are two types of deviations in the current execution cycle: sealing wrinkle deviation and label position deviation. Further calculation of the feature distance shows that the wrinkle deviation is moderate and the label position deviation is slight. Based on the built-in deviation-compensation mapping relationship, the model outputs a compensation value of reducing the heat-sealing temperature by 3 degrees Celsius for the sealing wrinkle deviation and a compensation value of shifting the label's Y-axis coordinate downwards by 0.2 mm for the label position deviation. The combination of these two values ​​forms the final packaging parameter compensation adjustment value.

[0055] In some embodiments of this application, the joint analysis model is constructed using a deep neural network. The input layer of the deep neural network includes multiple feature input branches corresponding to real-time packaging execution status data, post-packaging visual image data, upstream incoming material status data, and downstream logistics disturbance data, respectively. The multiple feature input branches are fused in the intermediate layer of the network to generate a joint feature vector.

[0056] In this embodiment, a deep neural network is a machine learning model inspired by the connection patterns of neurons in the human brain, composed of interconnected neurons at multiple layers. Each layer of neurons receives the output of the previous layer, performs weighted summation and nonlinear transformation, and then passes it to the next layer. Through this layer-by-layer processing, the deep neural network can automatically learn feature representations at different levels of abstraction from the raw input data. Deep neural networks have wide applications in image recognition, speech recognition, and natural language processing, and their deep structure gives them powerful feature learning capabilities. The input layer is the foremost layer in a deep neural network, responsible for receiving the raw data from the external input. The number of neurons in the input layer usually matches the dimensionality of the input data, with each neuron corresponding to a feature dimension of the input data. The input layer itself does not perform complex transformations on the data; it mainly serves to access and adapt the data, passing the external data to subsequent layers of the network for processing. Multiple feature input branches refer to setting relatively independent processing paths for different types of data at or near the input layer of the deep neural network. Because the four types of data—real-time packaging execution status data, post-packaging visual image data, upstream material arrival status data, and downstream logistics disturbance data—differ significantly in form and physical meaning, directly concatenating them using a uniform input layer might negatively impact the model's learning performance. By setting multiple feature input branches, each branch can be designed with a suitable network structure tailored to the characteristics of its corresponding data type. For example, for the post-packaging visual image data branch, a convolutional neural network can be used to process the image data, extracting visual features such as texture, edges, and shape. For the real-time packaging execution status data branch, a fully connected network or a one-dimensional convolutional network can be used to process the numerical time-series data, extracting the changing trends and statistical characteristics of parameters such as temperature and pressure. For upstream material arrival status data and downstream logistics disturbance data, corresponding branch structures can also be set according to their data characteristics. The intermediate layer of the network refers to the hidden layer located between the input and output layers. In a network structure with multiple feature input branches, data from different branches, after being processed through the preceding layers of their respective branches, converge at a certain intermediate layer. This convergence point is called the intermediate layer where feature fusion occurs. The number of intermediate layers and the number of neurons in each layer can be designed according to the complexity of the specific problem.

[0057] Feature fusion refers to the process of combining feature vectors from different branches in the middle layers of a network to form a unified joint feature representation. Feature fusion can be implemented in various ways. For example, feature vectors from different branches can be concatenated along a certain dimension to form a longer feature vector; weighted summation of feature vectors from different branches can also be used; or more complex attention mechanisms can be employed to allow the network to learn the importance of features from different branches and adaptively fuse them. Through feature fusion, information originally scattered across different branches, each describing a specific aspect of the packaging process, is integrated, making it possible for subsequent network layers to make decisions based on all the information simultaneously.

[0058] Understandably, the timing and method of feature fusion affect model performance. If fusion is too early, features from different branches haven't been fully extracted, making it difficult to establish effective relationships during fusion; if fusion is too late, features from each branch are already highly abstracted, potentially losing detailed information about interactions between branches. This technical solution sets the fusion point in the intermediate layer, allowing each branch to extract features from its own input data to a certain extent while preserving space for information interaction at an appropriate level of abstraction. The joint feature vector is a feature representation generated after feature fusion, incorporating information from multiple sources. Unlike features obtained by simply concatenating multiple-source joint analysis datasets, the joint feature vector here is extracted separately by each branch of the deep neural network and fused in the intermediate layer. It contains high-level semantic features extracted by each branch and the interaction information between these features, providing a more comprehensive reflection of the overall state of the current execution cycle's packaging process.

[0059] This application's technical solution constructs a joint analysis model specifically designed for multi-source heterogeneous data by setting up a deep neural network structure with multiple feature input branches and intermediate layer feature fusion. Different branches process different types of data, and each branch can adopt a network structure suitable for its data type. This solves the problem of the model's difficulty in effectively learning when image, numerical, and time-series data are directly mixed together. The visual image data branch can extract spatial structural features such as texture and edges through convolutional layers; the execution state data branch can capture the specific values ​​of parameters such as temperature and pressure and their interrelationships through fully connected layers; and the incoming material data and perturbation data branches can also be designed with appropriate processing methods according to their respective characteristics. Multiple branches perform feature fusion in the intermediate layers of the network, realizing the interaction of information from different sources at an appropriate level of abstraction. Before fusion, each branch has already performed a certain degree of feature extraction on the original data, transforming the original data into intermediate features with certain semantic meaning. The fusion process in the intermediate layer allows branch features from the image to be correlated with branch features from the execution state, such as linking the texture features of the sealing image with the heat-sealing temperature feature. Branch features from upstream material can also interact with branch features from downstream disturbances; for example, the features of filling weight and conveyor vibration, after fusion, jointly influence the judgment of labeling position deviation. This fusion method enables the model to learn the coupling relationships between different factors, providing structural support for accurately outputting the joint feature vector. The resulting joint feature vector, due to the fusion of features extracted from multiple branches in the intermediate layer, has higher information richness than any single-branch feature, and also higher than features obtained by simply concatenating them at the input layer. This joint feature vector serves as the basic input for subsequent layers and can be used for tasks such as deviation type identification, deviation degree calculation, and compensation adjustment value output. Since the joint feature vector itself already contains fusion information from multiple sources, subsequent tasks can make decisions based on more comprehensive information.

[0060] Taking a cigarette packaging production line as an example, the constructed joint analysis model adopts a deep neural network structure with four branches. The first branch processes the visual image data after packaging, using a structure of two convolutional layers and one pooling layer to extract texture features and label edge features from the sealing area image. The second branch processes real-time packaging execution status data, using a three-layer fully connected network, inputting values ​​such as heat sealing temperature, heat sealing pressure, and labeling position coordinates. The third branch processes upstream material arrival status data, using a two-layer fully connected network, inputting real-time filling weight and filling density. The fourth branch processes downstream logistics disturbance data, using a one-layer fully connected network, inputting conveying speed and vibration amplitude. After each of the four branches extracts features, feature concatenation and fusion are performed in the middle layer of the third layer of the network, concatenating the 64-dimensional features from the image branch, the 32-dimensional features from the execution status branch, the 16-dimensional features from the material arrival branch, and the 8-dimensional features from the disturbance branch into a 120-dimensional joint feature vector. This joint feature vector is then input into subsequent fully connected layers to output deviation type classification results and compensation adjustment value regression results. Through this branch fusion structure, the model can associate the subtle wrinkles in the sealing image with the specific value of the heat sealing temperature and the fluctuation of the filling weight. When outputting the compensation adjustment value, these factors are taken into account. For example, when the image shows sealing wrinkles and the incoming material is overfilled, the model will output a larger cooling compensation value than when there are only wrinkles.

[0061] In some embodiments of this application, the acquisition of upstream incoming material status data and downstream logistics disturbance data corresponding to the finished packaging includes: acquiring the real-time filling weight and filling density when filling to form the contents of the finished packaging, as upstream incoming material status data; and acquiring the real-time operating speed and vibration amplitude of the conveying mechanism during the process of the finished packaging being transported to the next workstation after packaging, as downstream logistics disturbance data.

[0062] It is understood that in this embodiment, real-time filling weight refers to the material mass value measured in real time by a weighing sensor or flow meter when injecting contents into the packaging bag or container during the upstream filling process of finished packaging. Filling weight is typically expressed in grams or kilograms, reflecting the amount of product contained in each individual packaging unit. During continuous production, filling weight may fluctuate slightly due to factors such as pressure fluctuations in the feeding system and changes in material flowability. For example, in a cigarette box packaging production line, the real-time filling weight when filling cigarette sticks into each pre-made bag can be different values ​​such as 240 grams, 242 grams, or 238 grams. Filling density refers to the mass per unit volume of the filled material in its filled state, typically expressed in grams per cubic centimeter or kilograms per cubic meter. Filling density is affected by factors such as the physical properties of the material itself, temperature, and degree of compaction. For the same product, the filling density may vary between different batches or at different times. Changes in filling density affect the pressure distribution within the packaging bag and the space occupied by the material, thereby affecting the stress state of the sealing area and the packaging appearance. For example, the packing density of the cigarette packs may vary due to changes in the moisture content of the tobacco or the degree of compaction. Real-time operating speed refers to the actual speed at which the conveyor belt, conveyor rollers, or chain, etc., move from the finished package to the next workstation after the packaging process is complete. Operating speed is typically expressed in meters per second or meters per minute, and its stability affects the time and positional accuracy of the package's arrival at subsequent workstations. For example, the real-time operating speed of the conveyor belt may be set to 0.5 meters per second, but due to fluctuations in motor speed or load, the actual speed may fluctuate between 0.48 meters per second and 0.52 meters per second.

[0063] Vibration amplitude refers to the intensity of mechanical vibration generated by the conveyor during operation, typically expressed as displacement amplitude (mm), velocity amplitude (mm / s), or acceleration amplitude (m / s²). Vibration amplitude can be caused by the conveyor's own operating conditions, such as motor imbalance or bearing wear, or by external factors such as impacts from nearby equipment. Large vibration amplitudes can cause slight shifts in the position of packages during transport, especially in labeling processes requiring precise positioning, where such shifts can affect the label's placement accuracy. For example, a vibration acceleration sensor on the conveyor platform might detect vibration amplitudes ranging from 0.1 m / s² to 0.3 m / s². It is understood that upstream material status data, in addition to real-time filling weight and density, may include other relevant parameters as needed, such as filling temperature, filling pressure, and material viscosity. Downstream material disturbance data, besides real-time operating speed and vibration amplitude, may include the number of start-stop cycles of the conveyor, the timing of steering mechanism movements, and accumulation pressure on the conveyor line.

[0064] This application's technical solution clarifies the actual collection content and physical meaning of upstream incoming material status data and downstream logistics disturbance data by specifying them concretely. Real-time filling weight and filling density, as upstream incoming material status data, reflect the quality and density of the contents of each packaging unit. These two parameters jointly determine the internal support force exerted by the contents on the inner wall of the packaging bag. Greater filling weight and higher filling density result in increased pressure inside the bag, altering the stress state of the material in the sealing area and affecting the fusion effect and sealing appearance during heat sealing. Real-time operating speed and vibration amplitude, as downstream logistics disturbance data, reflect the motion environment and mechanical disturbances experienced by the packaging during transport. Operating speed determines the moment the packaging arrives at the labeling station; speed fluctuations can cause deviations in the packaging's position over time. Vibration amplitude directly affects the positional stability of the packaging during transport; excessive vibration can cause slight displacement or shaking, leading to positioning errors at the labeling moment. These two sets of data, along with packaging execution status data and post-packaging visual image data, constitute a complete description of the packaging process information. Upstream incoming material status describes the initial state of the package upon entering the packaging process, while downstream material flow disturbances describe the environmental conditions after the package leaves the packaging process. By incorporating these two sets of data into a joint analysis, the model can identify whether packaging quality deviations stem from internal factors caused by changes in the characteristics of the contents, external factors caused by the influence of the transport environment, or the result of multiple factors working together. This multi-dimensional information input allows subsequent compensation adjustment value calculations to more accurately target the root cause of the deviation, rather than merely addressing superficial phenomena.

[0065] Taking a cigarette packaging production line as an example, in the upstream filling process, the weighing sensor measured the real-time filling weight of the cigarette pack at 245 grams, while the density meter measured the filling density at 0.48 grams per cubic centimeter. These two values ​​are slightly higher than the standard values ​​of 240 grams and 0.45 grams per cubic centimeter, respectively. In the downstream conveying process, the encoder measured the real-time conveyor belt speed at 0.48 meters per second, slightly lower than the standard speed of 0.5 meters per second. The accelerometer recorded a vibration amplitude of 0.25 millimeters during the conveying process, slightly higher than the usual 0.1 millimeters. These data, along with information such as the heat-sealing temperature of 142 degrees Celsius and the presence of wrinkles in the sealing image, were input into a joint analysis model. After analysis, the model determined that the excessive filling weight and high filling density led to increased pressure inside the bag, which was the main reason for wrinkles in the sealing area. Therefore, a compensation adjustment value of 3 degrees Celsius was output to reduce the heat-sealing temperature. At the same time, the slow running speed and excessive vibration amplitude may cause a time delay and positional deviation in the arrival of the package at the labeling station. Therefore, a compensation adjustment value of 0.1 seconds was output to delay the labeling timing.

[0066] In some embodiments of this application, real-time packaging execution status data, post-packaging visual image data, upstream material arrival status data, and downstream logistics disturbance data are aligned in the time dimension, including: taking the end time of the current execution cycle as a benchmark, tracing back to the upstream material filling time corresponding to the finished product packaging, associating the upstream material arrival status data generated at the upstream material filling time with the end time; and binding the real-time packaging execution status data and post-packaging visual image data corresponding to the end time with the associated upstream material arrival status data using timestamps.

[0067] It is understood that, in this application, the end time of the current execution cycle refers to the point in time when a finished package completes its packaging process. On the packaging production line, each finished package corresponds to an execution cycle, which typically begins when the packaging material or the item to be packaged enters the packaging station and ends when the main actions such as sealing and labeling are completed. The end time can be selected as the moment when the heat-sealing mechanism opens and releases the packaging bag, or it can be selected as the moment when the labeling mechanism completes the labeling action, depending on the actual situation of the production process. This moment serves as a reference point for time alignment and is used to correlate data from other processes. The upstream material filling moment refers to the specific point in time when the contents corresponding to the current finished package are filled into the packaging bag or packaging container. Since the filling process usually occurs before the packaging process, there is a time difference between the two processes, and the filling moment is earlier than the end time of the current execution cycle. The filling moment can be obtained through the event log recorded by the production line's control system. Each packaging unit is assigned a unique identifier or recorded with a precise timestamp when it is filled, for subsequent traceability. For example, in a cigarette packaging production line, the filling time of a certain carton might be 14:23:45.200, while the sealing time is 14:23:48.500, a difference of 3.3 seconds. Forward tracing refers to the process of finding the upstream material filling time corresponding to the current finished packaging, starting from the end time of the current execution cycle and moving in the reverse direction of the production process. Since the production line operates continuously, different packaging units flow sequentially along the line, requiring the establishment of a correspondence between upstream and downstream processes through production cycle time, sensor signals, or product identification codes. Forward tracing can be based on a fixed time offset; if the transport time from the filling station to the packaging station is fixed at T seconds, then the filling time of the current packaging is equal to the end time minus T seconds. Alternatively, it can be based on a product tracking system, where each packaging unit carries a QR code or RFID tag, and a reader records the accurate time it passes through each station. Association refers to establishing a correspondence between the upstream material status data generated at the upstream material filling time and the end time of the current execution cycle. The upstream incoming material status data was originally recorded and stored according to the filling time. Through correlation operations, this data is marked as belonging to the finished package of the current execution cycle, thus establishing a connection with other data of the same package. The correlated upstream incoming material status data can be regarded as part of the current package data and used for subsequent joint analysis.

[0068] Timestamp binding refers to the process of uniformly marking multi-source data corresponding to the same finished product packaging along the time dimension. Specifically, it involves assigning the same timestamp identifier to real-time packaging execution status data and post-packaging visual image data corresponding to the end time, or incorporating them into the same data record, along with the upstream incoming material status data after association processing. Through timestamp binding, multi-source data that were originally scattered across different data tables and recorded at different times are integrated into a complete data set describing the entire process of the packaging unit from upstream incoming materials to packaging completion.

[0069] Understandably, the collection time of downstream logistics disturbance data is usually later than the end time of the current execution cycle, because logistics disturbances occur during the transportation process after packaging is completed. For the time alignment of downstream logistics disturbance data, a similar but reversed approach can be adopted, that is, taking the end time of the current execution cycle as the benchmark, tracing backward to the moment when the package passes through key points such as the labeling and inspection station during transportation, and associating the logistics disturbance data collected at that moment with the end time and binding it with a timestamp.

[0070] The technical solution in this embodiment constructs a complete time alignment framework by using the end time of the current execution cycle as a benchmark, tracing back to the upstream material filling time and associating it, and then binding all data with timestamps. This framework solves the problem that multi-source data cannot be directly used for joint analysis due to different collection times. The forward tracing and association operations ensure that the upstream material status data can be accurately attributed to the corresponding finished packaging. Since the filling process precedes the packaging process, if the data is not traced and associated, directly arranging it in chronological order will lead to misalignment between upstream data and subsequent packaging, and the material status of packaging A may be mistakenly identified as the material status of packaging B during analysis. By tracing based on a fixed time offset or a product tracking system, a correct correspondence can be established. The timestamp binding operation integrates all data belonging to the same packaging together to form a complete data record. This record includes the material status of the packaging, the packaging execution status, the visual image after packaging, and the disturbance information during downstream transportation. The bound data can be directly input into the joint analysis model for processing. The model can analyze based on the complete information of the same packaging, rather than based on scattered data that are close in time but may belong to different packaging.

[0071] The technical benefit of this alignment method is that it ensures subsequent joint analysis is based on accurate causal relationships. During the learning process, the joint analysis model can establish a true mapping of how upstream materials affect packaging execution, how packaging execution affects packaging appearance, and how downstream disturbances affect subsequent processes. If the data alignment is inaccurate, the model may learn incorrect associations, leading to invalid compensation adjustment values ​​in the output.

[0072] Taking a cigarette box packaging production line as an example, the control system records the end time of the current execution cycle as 10:25:30:500 milliseconds. At this time, real-time execution status data such as heat sealing temperature of 142 degrees Celsius and labeling position coordinates X50.2Y79.8 are collected. Simultaneously, the vision system captures an image of the sealed package. Based on the production line setting that the conveying time from the filling station to the packaging station is fixed at 2.5 seconds, the system traces back to calculate the upstream material filling time as 10:25:28:000 milliseconds. From the filling process database, the system extracts the real-time filling weight of 245 grams and the filling density of 0.48 grams per cubic centimeter recorded at this time. At the same time, based on the conveying time after packaging is completed, the system traces back to the time when the package passes the downstream vibration monitoring point as 10:25:32:000 milliseconds, extracting the conveying speed of 0.48 meters per second and the vibration amplitude of 0.25 millimeters recorded at this time. The time alignment module binds all data and generates a complete record containing filling time data, packaging time data, and delivery time data. The timestamp is uniformly marked as the packaging batch number corresponding to 10:25:30:500 milliseconds.

[0073] In some embodiments of this application, the real-time packaging execution status data includes at least the heat sealing temperature, heat sealing pressure, and labeling position coordinates.

[0074] In this embodiment, heat sealing temperature refers to the instantaneous heating temperature of the sealing area when the heat sealing element of the heat sealing mechanism, such as a heat sealing knife or heat sealing rod, comes into contact with the packaging material during the packaging sealing process. Heat sealing temperature is usually expressed in degrees Celsius and is measured in real time by a temperature sensor, such as a thermocouple or resistance temperature detector (RTD), installed inside the heat sealing element. Heat sealing temperature is one of the key process parameters affecting the quality of packaging sealing. Excessive temperature may cause the packaging material to overheat and melt through or the sealing edges to become brittle, while excessively low temperature may result in a weak seal or inadequate sealing. For example, in the heat sealing process of BOPP film for cigarette packaging, the heat sealing temperature can be set between 140 and 150 degrees Celsius. Heat sealing pressure refers to the pressure per unit area applied to the heat sealing area by the heat sealing mechanism through drive elements such as pressure cylinders, servo motors, or mechanical cams during the packaging sealing process. Heat sealing pressure is usually expressed in megapascals (MPa) or kilopascals (kPa) and can be measured in real time by a pressure sensor installed on the pressure mechanism, or calculated based on the output force of the drive element and the heat sealing area. The heat-sealing pressure directly affects the tightness of the packaging material's adhesion in its molten state. Insufficient pressure may lead to insufficient seal strength, while excessive pressure may cause material deformation or molten material overflow. For example, the heat-sealing pressure can be set to 0.3 MPa, but in actual operation, it may vary between 0.28 MPa and 0.32 MPa due to air pressure fluctuations. Labeling position coordinates refer to the coordinate values ​​of the label's actual position on the packaging body relative to a preset reference point when the labeling mechanism releases and attaches the label to the packaging surface during the labeling process. Labeling position coordinates are usually measured in millimeters and represented using a two-dimensional coordinate system such as the horizontal X-axis and vertical Y-axis, or a three-dimensional coordinate system. The coordinate values ​​can be calculated using an encoder within the labeling mechanism combined with mechanical transmission, or confirmed by feedback from a post-labeling visual inspection system. The labeling position coordinates directly reflect the accuracy of label application; excessive coordinate deviation can cause the label to be skewed, tilted too high or too low, or exceed the allowable range. For example, the target coordinates of the upper left corner of the label on the packaging surface may be 50 mm on the X-axis and 80 mm on the Y-axis. In actual labeling, the coordinates may become 50.5 mm on the X-axis and 79.2 mm on the Y-axis due to positioning deviation.

[0075] Understandably, in addition to the three parameters mentioned above, real-time packaging execution status data can also include other relevant parameters depending on the specific needs of the packaging process. For example, for packaging requiring inflation protection, inflation pressure or inflation volume can be included; for packaging using ultrasonic sealing, ultrasonic amplitude or duration can be included; for packaging requiring printing of information such as production date, printhead temperature or printing pressure can be included. Heat sealing temperature, heat sealing pressure, and labeling position coordinates are the three most basic and important control parameters in the flexible packaging heat sealing and labeling process, and their real-time values ​​directly reflect the current working status of the packaging execution mechanism.

[0076] There is a certain correlation among these three parameters. Heat sealing temperature and heat sealing pressure work together on the sealing area, and their coordination determines the quality of the seal. When the temperature is too high, appropriately reducing the pressure can prevent excessive melting of the material; when the temperature is too low, appropriately increasing the pressure can improve the seal strength. Although the labeling position coordinates are independent of the heat sealing process, labeling accuracy may be affected by preceding heat sealing steps. For example, excessively high heat sealing temperatures can cause film shrinkage and deformation, changing the dimensions of the packaging bag and thus affecting the accuracy of the labeling position. Using these three parameters as core components of real-time packaging execution status data can provide status information reflecting key aspects of the packaging execution process for subsequent joint analysis.

[0077] In some embodiments of this application, the visual image data after packaging includes at least a texture feature map of the sealed area, a measurement of the sealed width, and the coordinates of the corner positions of the label.

[0078] It is understood that, in this embodiment, the texture feature map of the sealing area refers to image feature information extracted from the visual image after packaging, reflecting the surface texture and texture distribution of the packaging sealing area. During the heat sealing process, the material in the packaging sealing area melts under heat and then cools and solidifies, forming a specific texture morphology. Different heat sealing temperatures, pressures, and time conditions will produce different texture features. For example, when the temperature is moderate, the sealing texture is clear and uniform; when the temperature is too low, the sealing may have an incompletely fused, grainy feel; and when the temperature is too high, the sealing may have over-melted, smooth, or scorched areas. The texture feature map can be obtained through texture analysis methods in image processing technology, such as using existing methods like gray-level co-occurrence matrix, Gabor filter, or local binary mode (LBP) to extract texture features. It can also be automatically extracted and output as a feature map by the convolutional layer in a deep neural network. The texture feature map is presented in the form of an image or feature matrix, preserving the spatial distribution information of the texture and providing an intuitive basis for judging the sealing quality. The sealing width measurement value refers to the numerical result obtained after measuring the width dimension of the packaging sealing area. The sealing width typically refers to the lateral dimension of the heat-sealed area, i.e., the distance from the sealing edge to the junction of the seal and the bag body, usually measured in millimeters. Sealing width is a crucial quantitative indicator of sealing quality. A width that is too narrow may indicate insufficient heat-sealing pressure or heating time, while a width that is too wide may indicate excessive pressure or temperature leading to excessive material deformation. The sealing width measurement can be obtained by locating the sealing boundary using an edge detection algorithm in a visual image, calculating the pixel distance, and converting it into the actual physical size. For example, the standard sealing width for a certain size cigarette box packaging is 8 millimeters; an actual measurement of 7.5 millimeters or 8.5 millimeters may indicate a deviation. The corner position coordinates of a label refer to the numerical values ​​of the corner points of the label attached to the packaging surface in the image coordinate system or the packaging's physical coordinate system. Labels are typically rectangular or square with four corners, each represented by two-dimensional coordinates such as x and y coordinates. By detecting the corner position coordinates of the label, the actual attachment position and orientation of the label on the packaging surface can be determined, and whether it matches the preset position. The coordinates of the label corners can be obtained using corner detection algorithms in image processing or object detection models based on deep learning. For example, for a label on the front of a cigarette pack, the coordinates of the top left corner of the standard label are set to 10 mm from the left edge and 15 mm from the top edge of the pack. The actual detected coordinate values ​​can reflect whether the label has horizontal, vertical, or rotational tilt.

[0079] It is understandable that, in addition to the three specific types mentioned above, the visual image data after packaging can also include other image-derived features as needed, such as the grayscale histogram of the sealing area, the tilt angle of the label area, and the detection results of scratches or stains on the packaging surface. These features together constitute a quantitative description of the packaging appearance quality.

[0080] The technical solution in this application embodiment concretizes the visual image data after packaging into a texture feature map of the sealing area, a measurement value of the sealing width, and the coordinates of the corner positions of the label, constructing a description system of appearance quality from macro to micro and from qualitative to quantitative. The texture feature map of the sealing area preserves the fine texture information of the sealing surface in image form, which can reflect the subtle differences in the fusion state of the material during the heat sealing process, such as whether the fusion is uniform, whether there are unfused particles, and whether there is overheating and scorching. This texture information is highly intuitive and rich in detail, making it suitable for judging the overall condition and abnormal patterns of the sealing quality. The measurement value of the sealing width quantifies the key indicator of the sealing size in specific numerical form, providing an objective basis for direct comparison with process standards. Small changes in the width value may be directly related to fluctuations in heat sealing temperature and pressure. Combining the width measurement value with the texture feature map can mutually verify the judgment of the sealing quality. For example, if the texture feature map shows a blurry texture in the sealing area and an excessively wide seal, it may indicate that the heat-sealing temperature is too high; if the texture feature map shows a rough texture in the sealing area and an excessively narrow seal, it may indicate that the heat-sealing pressure is insufficient. The corner coordinates of the label precisely pinpoint its actual attachment position on the packaging surface, providing a quantitative standard for judging labeling accuracy. The center position offset and rotation angle of the label can be calculated using the four corner coordinates, comprehensively evaluating the labeling quality. Label position deviation may originate from positioning errors in the labeling mechanism itself, changes in the dimensions of the packaging bag during the previous heat-sealing process, or displacement of the packaging body during transport.

[0081] These three features describe the packaging appearance quality from three different perspectives: seal texture, seal size, and label position accuracy. They form a complementary and mutually reinforcing information system. The texture feature map provides intuitive visual information, suitable for identifying abnormal patterns; the seal width measurement provides quantitative dimensional information, suitable for precise comparison with process standards; and the label corner coordinates provide precise positioning information, suitable for judging labeling deviations. All three are input into the joint analysis model, enabling the model to comprehensively utilize visual details, dimensional values, and position coordinates for a comprehensive judgment, laying the foundation for subsequently outputting accurate compensation and adjustment values.

[0082] In some embodiments of this application, before inputting the multi-source joint analysis dataset into the pre-built joint analysis model, the method further includes: acquiring historical real-time packaging execution status data, historical post-packaging visual image data, historical upstream incoming material status data, historical downstream logistics disturbance data, and their corresponding historical compensation adjustment values ​​within a historical production cycle; using the historical real-time packaging execution status data, historical post-packaging visual image data, historical upstream incoming material status data, and historical downstream logistics disturbance data as training samples, and using the corresponding historical compensation adjustment values ​​as supervision labels, training the initial neural network model to obtain the joint analysis model.

[0083] Understandably, in this embodiment, the historical production cycle refers to a production operation period that has been completed before the current moment. The historical production cycle can include production records over the past few hours, days, weeks, or even longer periods, with the specific time span depending on the operational stability of the production line and the amount of data accumulated. Each historical production cycle corresponds to a complete packaging execution process, encompassing the entire process from upstream filling to packaging execution and downstream conveying. Historical real-time packaging execution status data refers to the process parameter data collected and recorded in real-time by sensors on the packaging execution mechanism during past production cycles. This data has the same type and format as the real-time packaging execution status data collected in the current production cycle, including, for example, historical heat-sealing temperature, heat-sealing pressure, and labeling position coordinates, only in the past. Similarly, historical post-packaging visual image data, historical upstream incoming material status data, and historical downstream logistics disturbance data correspond to post-packaging visual images, upstream filling process data, and downstream conveying process data collected in past production cycles, respectively. Historical compensation adjustment values ​​refer to the compensation values ​​actually applied to packaging parameter adjustments during past production cycles. These compensation adjustment values ​​may originate from various sources. For example, they could be adjustments manually input by experienced operators based on observed packaging quality, adjustments calculated using other control methods, or optimal adjustments calibrated through comparative experiments during the experimental phase. Historical compensation adjustment values ​​record the parameter correction schemes actually implemented to improve packaging quality under specific historical operating conditions. Training samples refer to the input data used to train the machine learning model. In this technical solution, the training samples consist of historical real-time packaging execution status data, historical post-packaging visual image data, historical upstream incoming material status data, and historical downstream logistics disturbance data. These data, aligned in the time dimension, form a complete multi-source joint analysis dataset, which serves as the model's input. Supervision labels are the expected output values ​​corresponding to the training samples in machine learning, used to guide the model in learning the mapping relationship from input to output. In this technical solution, supervision labels are historical compensation adjustment values, i.e., the packaging parameter compensation adjustment values ​​actually used in the historical production cycle corresponding to each training sample. By allowing the model to learn the correspondence between input multi-source data and output compensation adjustment values, the model can gradually master the pattern of what adjustment amount should be output under different operating conditions. An initial neural network model refers to a neural network structure that has not been trained or has only undergone preliminary initialization. This model has a predefined network architecture, such as the deep neural network structure with multiple feature input branches described earlier, but the connection weights within the network are randomly initialized or initialized based on a certain preset method, and have not yet acquired knowledge from the data through learning. The training process involves continuously adjusting these weights to gradually bring the model's output closer to the supervised label.Understandably, during training, historical data needs to be divided into training, validation, and test sets to evaluate the model's training performance and generalization ability. Training can be performed using batch training, where a batch of samples is input each time to calculate the loss function value, and the network weights are updated through backpropagation. This process is repeated multiple times until the model converges.

[0084] The technical solution in this application establishes a foundation for the joint analysis model to learn knowledge from data by constructing a historical dataset and using it for model training. Historical real-time packaging execution status data, historical post-packaging visual image data, historical upstream incoming material status data, and historical downstream logistics disturbance data serve as training samples, comprehensively recording all-round information from upstream to downstream and from execution to result in the past production process. These samples cover various combinations of operating conditions that may occur during the production process, including normal stable operating conditions, operating conditions with slight fluctuations, and operating conditions with significant deviations, providing rich material for the model to learn the patterns under different conditions. Historical compensation adjustment values ​​serve as supervision labels, providing guidance for the model's learning. These compensation adjustment values ​​represent parameter correction schemes considered effective under historical operating conditions, and whether derived from human experience or experimental calibration, they contain knowledge accumulated in actual operation. By learning the correspondence between input data and output adjustment values, the model gradually masters what parameter compensation should be applied to improve packaging appearance quality under different upstream incoming material states, different execution states, and different downstream disturbance conditions. The correspondence between training samples and supervision labels is the core of model learning. Each training sample corresponds to multi-source data from a historical production cycle, and each supervision label corresponds to the actual compensation adjustment value used in that cycle. Through numerous such correspondences, the model learns the mapping pattern between the implicit state features in the multi-source data and the appropriate adjustment amount. After sufficient training, when the model receives new multi-source data for the current cycle, it can output a compensation adjustment value suitable for the current operating conditions based on the learned patterns. This method of training the model based on historical data allows it to inherit the experiential knowledge accumulated in historical production, including human experience and experimental results. Compared to relying entirely on manually written adjustment logic, this method can capture the complex nonlinear relationships implicit in the data and handle situations involving multiple coupled factors. As production continues, new qualified adjustment cases can be continuously collected and added to the training set, enabling the model's knowledge to be continuously updated and optimized.

[0085] Taking a cigarette packaging production line as an example, the factory collected historical production data from the past three months as a training set. This included records of 10,000 historical production cycles. Each record contained execution status data such as the heat-sealing temperature of 142 degrees Celsius and the heat-sealing pressure of 0.31 MPa, sealing and label images, incoming material data such as the filling weight of 241 grams and the filling density of 0.46, disturbance data such as the conveyor speed of 0.49 meters per second and the vibration of 0.15 millimeters, and compensation adjustment values ​​manually entered by the operators at that time: a 2-degree Celsius decrease in heat-sealing temperature and a 0.2-mm downward shift in the label Y-axis. This historical data was input into an initial neural network model for training. The model continuously adjusted its internal weights so that, for the input historical multi-source data, the compensation adjustment values ​​output by the model gradually approached the compensation adjustment values ​​actually used in history. After training, when the production line experiences a current cycle heat sealing temperature of 142 degrees, a filling weight of 245 grams, and a vibration of 0.25 millimeters, the trained model, based on patterns learned from historical data, outputs a compensation adjustment value that reduces the heat sealing temperature by 3 degrees. This output is consistent with the adjustment experience under similar working conditions in historical data.

[0086] In some embodiments of this application, the historical compensation adjustment value is based on the historical best adjustment experience value marked manually or the best adjustment value obtained through closed-loop control test calibration.

[0087] It is understood that, in this embodiment, the manually annotated historical optimal adjustment experience value refers to the optimal parameter adjustment amount manually recorded or calibrated by experienced operators or process engineers based on long-term production practice and familiarity with equipment characteristics, for specific operating conditions encountered in historical production processes. In actual production, when operators observe a deviation trend in packaging quality, such as slight wrinkles in the seal or label misalignment, they will judge the parameters that need to be adjusted and the adjustment range based on experience. If the packaging quality improves after adjustment, reaching or approaching the ideal state, then the adjustment value used in this adjustment can be recorded as an optimal adjustment experience value under that operating condition. The manual annotation process can be done by clicking confirmation on the human-machine interface, or by using a dedicated annotation tool to associate the adjustment value with the corresponding production data. For example, an operator finds that when the filling weight is 3 grams too high and the heat sealing temperature is 142 degrees Celsius, wrinkles easily appear in the seal. After lowering the heat sealing temperature by 2 degrees Celsius, the wrinkles disappear, so he marks this adjustment value of lowering the temperature by 2 degrees Celsius as the optimal experience value for this operating condition. The optimal adjustment value obtained from closed-loop control test calibration refers to the optimal parameter adjustment amount determined through automated testing, systematic parameter scanning and effect evaluation under controlled conditions for a specific operating condition. Closed-loop control testing can be implemented in various ways. For example, during production line downtime maintenance, specific upstream material conditions and downstream disturbance conditions can be simulated. Packaging parameters are then changed in a step-by-step manner, recording visual image data after packaging under each parameter combination. Packaging quality is automatically evaluated through image analysis, and the parameter adjustment value that optimizes packaging quality indicators is ultimately selected as the calibration result. Alternatively, online calibration can be performed during production using small-scale trials. For example, under the current operating condition, the control system actively applies a small parameter adjustment trial, observing the trend of packaging quality changes. If the trial result is positive, the adjustment continues in the same direction; if negative, the adjustment reverses. This cycle of disturbance-observation-adjustment gradually approaches the optimal adjustment value. The optimal adjustment value obtained from closed-loop control test calibration does not rely on human experience but is derived through systematic testing and data evaluation, possessing repeatability and objectivity. Understandably, manually labeled empirical values ​​and optimal values ​​calibrated through closed-loop testing can complement each other. In the early stages of a production line or when historical data is scarce, manually labeled empirical values ​​can be primarily relied upon to construct the initial training sample library. As production data accumulates, manually labeled values ​​can be validated and corrected through closed-loop testing, or new calibration data can be generated entirely from closed-loop testing. Adjusted values ​​from both sources can serve as supervisory labels during the training of the joint analysis model, guiding the model to learn the mapping relationship from multi-source data to the optimal compensation adjustment value.

[0088] In some embodiments of this application, updating the target packaging parameters for the next execution cycle based on the packaging parameter compensation adjustment value can be achieved in the following way: First, obtain the initial target packaging parameters preset for the next execution cycle. These initial target packaging parameters are baseline values ​​predetermined based on production process standards or historical statistical data. Then, the packaging parameter compensation adjustment value output from the joint analysis model is superimposed with the initial target packaging parameters. The specific method of superposition calculation can be addition, that is, adding the values ​​of each corresponding parameter dimension separately. For example, adding the initial target heat sealing temperature with the heat sealing temperature compensation adjustment value, adding the initial target heat sealing pressure with the heat sealing pressure compensation adjustment value, and adding the initial target labeling position coordinates with the labeling position compensation adjustment value. After superposition calculation, a set of updated target packaging parameters is generated. This set of updated parameters includes the baseline process requirements and the deviation compensation requirements identified by the joint analysis results of multi-source data from the previous execution cycle. The updated target packaging parameters are then sent to the control system of the packaging execution mechanism as the control target for the actual operation of the next execution cycle.

[0089] The technical solution in this embodiment of the application achieves an organic combination of baseline process requirements and dynamic compensation needs by superimposing the initial target packaging parameters and the packaging parameter compensation adjustment values. The initial target packaging parameters are baseline values ​​predetermined based on production process standards or historical statistical data, reflecting the parameter settings required to achieve the required packaging quality under ideal operating conditions. These parameters provide a stable process foundation and traceable compliance basis for the production process. The packaging parameter compensation adjustment values ​​are deviation compensation needs identified by the joint analysis model based on the results of multi-source data joint analysis of the previous execution cycle. They include dynamic corrections to address factors such as upstream material fluctuations and downstream environmental disturbances. The two are superimposed through an additive operation, adding values ​​for each parameter dimension, such as heat-sealing temperature, heat-sealing pressure, and labeling position coordinates. This ensures that the updated target packaging parameters retain the stability of the original process baseline while incorporating predictive adjustments based on current operating condition trends.

[0090] The technical advantage of this superimposed update method lies in the stability and controllability of parameter adjustments. Directly replacing the original target parameters with compensation adjustment values ​​may lead to excessive adjustment magnitude or directional deviation. The superimposed method, however, ensures that each adjustment is an incremental correction based on the original baseline, helping to avoid process fluctuations caused by single large adjustments. Simultaneously, since the compensation adjustment values ​​are targeted corrections derived from multi-source data joint analysis, the superimposed parameters can more accurately match the current upstream material characteristics and downstream conveying environment. This allows actuators such as the heat-sealing mechanism and labeling mechanism to operate under more suitable conditions, reducing sealing defects or labeling misalignment caused by parameter mismatch. The updated parameters, calculated through superposition, are sent to the control system as the actual control target, completing a full closed loop from data acquisition and joint analysis to parameter adjustment, enabling packaging parameters to evolve smoothly in response to changes in production line conditions.

[0091] In some embodiments of this application, the method may further include a model incremental update process based on positive samples. After the next execution cycle ends, updated visual image data of the finished product packaging produced in that cycle is acquired through a visual acquisition device; the updated visual image data is input to a quality evaluation module, which performs comparative analysis based on preset appearance quality standards; the preset appearance quality standards may include indicators such as sealing integrity threshold, label position tolerance range, and allowable deviation of sealing width; if all indicators of the updated visual image data are within the preset standard range, the packaging quality of that execution cycle is determined to meet the requirements; at this time, the packaging parameter compensation adjustment value used in that execution cycle and the multi-source joint analysis dataset used to generate the compensation adjustment value are extracted and combined to form a positive sample; this positive sample is added to the training sample library of the joint analysis model for incremental training of the model.

[0092] This application constructs a self-optimizing closed-loop system by introducing a positive sample incremental update mechanism. After the next execution cycle, updated visual image data of the packaged product is first acquired through a visual acquisition device. This step serves as the input for quality feedback, providing an objective basis for subsequent evaluation. The quality evaluation module analyzes the image data based on preset appearance quality standards. These standards include specific quantitative indicators such as the seal integrity threshold, label position tolerance range, and allowable deviation of seal width, making the judgment of quality compliance clearly operable. When all indicators are within the standard range, it indicates that the packaging parameter compensation adjustment values ​​used in this cycle have achieved the expected effect, successfully meeting the packaging quality requirements. At this point, the compensation adjustment values ​​used in this cycle, along with the multi-source joint analysis dataset used to generate these adjustment values, are extracted and combined to form a positive sample. This positive sample records the complete mapping relationship from specific working condition input to successful adjustment output, serving as a positive verification of the joint analysis model's decision-making capability. Positive samples are added to the training sample library for incremental training of the model, enabling it to continuously learn from successful adjustment experiences, constantly reinforcing correct decision paths, and gradually improving its ability to output accurate compensation adjustment values ​​under new operating conditions. These four steps are interconnected: visual acquisition provides feedback data, quality assessment generates pass / fail judgments, positive samples construct successful cases, and incremental training optimizes the model. Through this mechanism, the joint analysis model can continuously adapt to the slow changes in production line conditions, seasonal fluctuations in material properties, and equipment aging and drift during actual production, maintaining long-term stable optimization results without frequent manual intervention and model redevelopment.

[0093] In some embodiments of this application, the finished product packaging can adopt a pre-made bag filling packaging form using flexible packaging materials; flexible packaging materials include plastic film, aluminum foil composite film, or composite materials of paper and plastic; pre-made bags refer to pre-formed and opened packaging bags, in which the contents are filled into the bag by a filling mechanism during the filling process, and then the bag enters the heat sealing process to complete the sealing; the packaging execution mechanism may include a film pulling mechanism, a heat sealing mechanism, and a labeling mechanism; the film pulling mechanism is used to pull the film material forward in a continuous packaging machine and control the bag length and positioning accuracy of each packaging unit; the heat sealing mechanism includes a transverse heat sealing assembly and a longitudinal heat sealing assembly, which, through the cooperation of a heating element and a pressurizing cylinder, applies heat and pressure to the bag opening at a set temperature to complete the fusion seal; the labeling mechanism is located after the heat sealing process or in parallel with the heat sealing process, and affixes the label to the designated position of the packaging bag.

[0094] The technical solutions in this application clearly define the specific form of finished product packaging and the composition of the packaging execution mechanism, implementing the aforementioned abstract method steps into the specific application scenario of flexible pre-made bag filling packaging. The characteristics of flexible packaging materials such as plastic film and aluminum foil composite film make them highly sensitive to parameters such as heat-sealing temperature, pressure, and time. Excessively high temperatures may cause the film to melt through, while excessively low temperatures will prevent effective sealing. This places specific requirements on the collection of packaging execution status data and the calculation of compensation adjustment values. The forming and filling process of the pre-made bag means that the filling weight and filling density in the upstream incoming material status data directly affect the pressure distribution inside the bag, thereby affecting the stress state of the heat-sealing area. This physical correlation closely links the upstream incoming material data with the packaging execution parameters. The film-pulling mechanism controls the bag length and positioning accuracy, and its operational stability affects the positional accuracy of each packaging unit at the heat-sealing and labeling stations. The heat-sealing mechanism includes transverse and longitudinal heat-sealing components, which complete the sealing through the cooperation of heating elements and pressure cylinders. Real-time data on heat-sealing temperature and pressure are important bases for judging the sealing quality. The labeling mechanism, operating after or concurrently with the heat-sealing process, affixes labels to designated positions. Its labeling accuracy is influenced by material shrinkage and deformation caused by the preceding heat-sealing process, as well as the operating speed and vibration amplitude of the downstream conveyor mechanism. These actuators have clear temporal relationships and physical couplings. Positioning errors during film stretching are transmitted to the heat-sealing process; temperature and pressure changes during heat sealing affect material dimensions and thus labeling positions; and vibrations in the conveyor mechanism interfere with the labeling timing. The method in this application collects real-time status data from these mechanisms during operation and combines it with upstream material data and downstream disturbance data for joint analysis. This allows the output compensation adjustment value to comprehensively consider the mutual influences between the mechanisms. For example, when conveyor vibration is detected causing labeling position deviation, the labeling trigger timing can be adjusted, or the heat-sealing temperature can be fine-tuned simultaneously to reduce material shrinkage and deformation, thereby achieving collaborative optimization of multiple actuators.

[0095] Please refer to Figure 2 A second aspect of this application provides a finished product packaging parameter optimization system based on joint data analysis. This system is used to implement the aforementioned finished product packaging parameter optimization method based on joint data analysis, including: The first data acquisition module 21 is used to acquire real-time packaging execution status data of the finished product packaging in the current execution cycle; The second data acquisition module 22 is used to collect the visual image data of the finished product packaging after the current execution cycle ends. The third data acquisition module 23 is used to acquire upstream incoming material status data and downstream logistics disturbance data corresponding to the finished product packaging. The data alignment module 24 is used to align real-time packaging execution status data, post-packaging visual image data, upstream material arrival status data, and downstream logistics disturbance data in the time dimension to generate a multi-source joint analysis dataset. The compensation value calculation module 25 is used to input the multi-source joint analysis dataset into the pre-built joint analysis model and output the packaging parameter compensation adjustment value for the next execution cycle; wherein, the joint analysis model is a model used to describe the relationship between packaging execution status, incoming material status, logistics disturbance and post-packaging appearance characterization; Parameter update module 26 is used to update the target packaging parameters for the next execution cycle based on the packaging parameter compensation adjustment value; The execution control module 27 is used to control the operation of the packaging execution mechanism based on the updated target packaging parameters in the next execution cycle.

[0096] In the above embodiments, the packaging parameter optimization process can be automated and made real-time through the collaborative work of each module.

[0097] See Figure 3 , Figure 3 This is a schematic block diagram of an electronic device provided according to an embodiment of this application. Figure 3 The electronic device 300 in this embodiment may include one or more processors 301, one or more input devices 302, one or more output devices 303, and one or more memories 304. The processors 301, input devices 302, output devices 303, and memories 304 communicate with each other via a communication bus 305. The memories 304 store computer programs, including program instructions. The processors 301 execute the program instructions stored in the memories 304. Specifically, the processors 301 are configured to invoke the program instructions to execute the aforementioned method for optimizing finished product packaging parameters based on joint data analysis.

[0098] It should be understood that, in the embodiments of this application, the processor 301 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0099] Input device 302 may include a touchpad, a fingerprint sensor (for collecting the user's fingerprint information and fingerprint orientation information), a microphone, etc., and output device 303 may include a display (LCD, etc.), a speaker, etc.

[0100] The memory 304 may include read-only memory and random access memory, and provides instructions and data to the processor 301. A portion of the memory 304 may also include non-volatile random access memory. For example, the memory 304 may also store device type information.

[0101] In specific implementations, the processor 301, input device 302, and output device 303 described in the embodiments of this application can execute the implementation methods described in any embodiment of the finished product packaging parameter optimization method based on data joint analysis provided in the embodiments of this application, or they can execute the implementation methods of the electronic devices described in the embodiments of this application, which will not be repeated here.

[0102] In another embodiment of this application, an electronic device is provided. The electronic device stores a computer program, which includes program instructions. When executed by a processor, the program instructions implement all or part of the processes in the above-described method for optimizing finished product packaging parameters based on data joint analysis. Alternatively, the computer program can instruct related hardware to complete the process. The computer program can be stored in an electronic device, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. Computer-readable media can include any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0103] The computer-readable storage medium can be an internal storage unit of the electronic device in any of the foregoing embodiments, such as a hard disk or memory of the electronic device. The computer-readable storage medium can also be an external storage device of the electronic device, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the electronic device. Furthermore, the computer-readable storage medium can include both internal and external storage units of the electronic device. The computer-readable storage medium is used to store computer programs and other programs and data required by the electronic device. The computer-readable storage medium can also be used to temporarily store data that has been output or will be output.

[0104] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this application.

[0105] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the electronic devices and units described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0106] In the several embodiments provided in this application, it should be understood that the disclosed electronic devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces or units, or it may be an electrical, mechanical, or other form of connection.

[0107] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiments of this application, depending on actual needs.

[0108] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0109] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for optimizing finished product packaging parameters based on joint data analysis, characterized in that, Includes the following steps: S1: Obtain real-time packaging execution status data of the finished product packaging within the current execution cycle; S2: After the current execution cycle ends, collect the visual image data of the finished product packaging after packaging; S3: Obtain upstream incoming material status data and downstream logistics disturbance data corresponding to the finished product packaging; S4: Align the real-time packaging execution status data, the post-packaging visual image data, the upstream incoming material status data, and the downstream logistics disturbance data in the time dimension to generate a multi-source joint analysis dataset; S5: Input the multi-source joint analysis dataset into the pre-built joint analysis model and output the packaging parameter compensation adjustment value for the next execution cycle; wherein, the joint analysis model is a model used to describe the correlation between packaging execution status, incoming material status, logistics disturbance and post-packaging appearance characterization; S6: Update the target packaging parameters for the next execution cycle based on the packaging parameter compensation adjustment value; S7: In the next execution cycle, control the operation of the packaging actuator based on the updated target packaging parameters.

2. The method according to claim 1, characterized in that, The step of inputting the multi-source joint analysis dataset into a pre-built joint analysis model and outputting packaging parameter compensation adjustment values ​​for the next execution cycle includes: extracting features from the multi-source joint analysis dataset through the joint analysis model to obtain a joint feature vector; matching the joint feature vector with a preset standard packaging feature template to determine the type and degree of packaging quality deviation in the current execution cycle. Based on the type and degree of deviation, calculate the packaging parameter compensation adjustment value to offset the deviation.

3. The method according to claim 2, characterized in that, The joint analysis model is constructed using a deep neural network. The input layer of the deep neural network includes multiple feature input branches corresponding to the real-time packaging execution status data, the post-packaging visual image data, the upstream incoming material status data, and the downstream logistics disturbance data, respectively. The multiple feature input branches are fused in the middle layer of the network to generate the joint feature vector.

4. The method according to claim 1, characterized in that, The acquisition of upstream incoming material status data and downstream logistics disturbance data corresponding to the finished packaging includes: acquiring the real-time filling weight and filling density when filling to form the contents of the finished packaging, as the upstream incoming material status data; and acquiring the real-time operating speed and vibration amplitude of the conveying mechanism during the process of the finished packaging being transported to the next workstation after packaging, as the downstream logistics disturbance data.

5. The method according to claim 1, characterized in that, Aligning the real-time packaging execution status data, the post-packaging visual image data, the upstream material arrival status data, and the downstream logistics disturbance data in the time dimension includes: taking the end time of the current execution cycle as a benchmark, tracing back to the upstream material filling time corresponding to the finished product packaging, and associating the upstream material arrival status data generated at the upstream material filling time with the end time; The real-time packaging execution status data and the post-packaging visual image data corresponding to the end time are timestamped and bound to the associated upstream incoming material status data.

6. The method according to claim 1, characterized in that, The real-time packaging execution status data includes at least the heat sealing temperature, heat sealing pressure, and labeling position coordinates.

7. The method according to claim 1, characterized in that, The post-packaging visual image data includes at least the texture feature map of the sealed area, the measurement value of the sealed width, and the coordinates of the corner positions of the label.

8. The method according to claim 1, characterized in that, Before inputting the multi-source joint analysis dataset into the pre-built joint analysis model, the method further includes: acquiring historical real-time packaging execution status data, historical post-packaging visual image data, historical upstream incoming material status data, historical downstream logistics disturbance data, and their corresponding historical compensation adjustment values ​​within a historical production cycle; using the historical real-time packaging execution status data, historical post-packaging visual image data, historical upstream incoming material status data, and historical downstream logistics disturbance data as training samples, and using the corresponding historical compensation adjustment values ​​as supervision labels, training an initial neural network model to obtain the joint analysis model.

9. The method according to claim 8, characterized in that, The historical compensation adjustment value is based on the historical best adjustment experience value marked manually or the best adjustment value obtained through closed-loop control test calibration.

10. A finished product packaging parameter optimization system based on joint data analysis, used to execute the finished product packaging parameter optimization method based on joint data analysis as described in any one of claims 1-9, characterized in that, include: The first data acquisition module is used to acquire real-time packaging execution status data of finished product packaging within the current execution cycle; The second data acquisition module is used to collect the finished product packaging visual image data after the current execution cycle ends; The third data acquisition module is used to acquire upstream material status data and downstream logistics disturbance data corresponding to the finished product packaging. The data alignment module is used to align the real-time packaging execution status data, the post-packaging visual image data, the upstream material arrival status data, and the downstream logistics disturbance data in the time dimension to generate a multi-source joint analysis dataset. The compensation value calculation module is used to input the multi-source joint analysis dataset into a pre-built joint analysis model and output the packaging parameter compensation adjustment value for the next execution cycle; wherein, the joint analysis model is a model used to describe the correlation between packaging execution status, incoming material status, logistics disturbance and post-packaging appearance characterization; The parameter update module is used to update the target packaging parameters for the next execution cycle based on the packaging parameter compensation adjustment value. The execution control module is used to control the operation of the packaging execution mechanism based on the updated target packaging parameters in the next execution cycle.

Images